Purpose/Significance To address the hallucination problem in discharge summary generation by using large language models, and to enhance the generative ability and contextual consistency of large language models. Method/Process The study constructs a high-quality, multi-level medical instruction dataset and employs a staged training-based instruction fine-tuning strategy to guide the model in learning tasks from simple to complex. A data replay and mixed training mechanism is introduced during the fine-tuning process to ensure that the large language model retains and utilizes existing knowledge when tackling new tasks. Result/Conclusion Experimental results show that the method proposed in this paper significantly reduces the occurrence of hallucinations in large language model generation and improves the accuracy and reliability of medical text generation. The superiority of this method lies in the effective integration of curriculum learning theory and the replay mechanism, which not only enhances the model’s adaptability to complex tasks, but also ensures the professionalism of the generated content, while showing high practicality and reliability.